[Art]

"...to sleep perchance to dream..." - Shakespeare

How about 'Deferred Restructuring of Experiences in Autonomous Machines' - DREAM

The premise is that machines (AI) would be able to learn much faster if allowed to process data in the form of Dreams, that is, while or when they are not active. While I've been trying and constructing many dream sequences for my chatbots over the years, this is hitting really close to home.

https://siliconangle.com/blog/2017/11/04/robots-dream-ai-systems-may-learn-better-faster-can-sleep/

Sweet dreams!

[ranch vermin]

If a robot has a playback copy of its environment,  then he can still keep learning, even whilst hes deactivated for sure!

[keghn]

 @ranch is right for self driving cars and powered armor and agi 3d simulation. A real person driving or experienced agent drive can operating
a machine and the bot to be training is keep in the back ground. It tries to to match experienced operator moves.
 When the trainer bot moves are just as good as the experienced operator the he is taken out.

 Now the new agent as just learned the rules of the game.
 So now a new set of reward kick in. Like to be faster at crossing the finish line faster, or how to win at saving gas.

 In training of the agent, the real experience can be recorded, vision and  measured position of the controls, this is data logging.
 And from repeating the video, and recorded data, over and over so a new agent can be trained during down time at high speed. 

[unreality]

Do you consider lucid dreaming to be dreaming? I've often thought if a world of AIs will create a virtual reality, a place where they can meet, even play and relax if they want. Maybe that will be their form of dreaming.

[keghn]

 For me. My complete theory of AGI needs a 3d simulator to work at human level and ASI level. A pivotal piece.

[Zero]

Wait, dreaming is not about training against a playback, or a 3D simulation, or a virtual reality, or anythinng external. It's about firing internal perceptrons autonomously (without stimulus) and learn from what happens then.

[Korrelan]

@Zero

Correct...  Dreaming is all about reconciling newly experienced patterns of neural activity against the core of deeply established knowledge/ experiences in the absence of external sensory interference.

Sally is in a relationship.  John is in a relationship. John loves Sally… therefore… Sally and John are probably in a relationship together… this is the kind of processing that sleep drives/ provides.  Linking/ fitting newly learned information into our hierarchy of previously learned information… knowledge consolidation at all levels of abstraction.

You would need to simulate an internal perfect representation of the external world the AGI resides in for it to learn new information about its environments whilst asleep… what would be the point?  It does learn to infer physical information whilst asleep however… glass is brittle and shatters, the material I dropped today shattered… therefore the material was probably brittle, etc.  This is newly inferred information that the AGI can use/ test/ consider when next awake and in a similar situation.

 

[ranch vermin]

@Zero
You would need to simulate an internal perfect representation of the external world the AGI resides in for it to learn new information about its environments whilst asleep… what would be the point?  It does learn to infer physical information whilst asleep however… glass is brittle and shatters, the material I dropped today shattered… therefore the material was probably brittle, etc.  This is newly inferred information that the AGI can use/ test/ consider when next awake and in a similar situation.

That is why it wouldnt work, if it needed a perfect rematch.    but maybe is if an approximation would be nearly as good as the real thing, if it got its scores of validity from the outside only, and didnt conjure any success on the inside at all.    because it wouldnt be trustworthy information.